An oversampling-integrated deep neural network for imbalanced credit card fraud detection
Rinku1, Ashutosh Kumar Dubey1 and Sushil Kumar Narang2
Chitkara University Institute of Engineering and Technology,Chitkara University,Punjab,India2
Corresponding Author : Rinku
Recieved : 09-September-2025; Revised : 23-December-2025; Accepted : 25-December-2025
Abstract
Credit card fraud detection remains challenging due to the need for the highly reliable identification of fraudulent transactions. In this paper an oversampling-integrated deep neural network (DNN) framework was proposed to improve fraud detection. Experiments were conducted on European credit card transaction dataset. Three oversampling approaches (synthetic minority over-sampling technique (SMOTE), adaptive synthetic sampling (ADASYN), and borderline-SMOTE) were used for handling class imbalance. Then, a multi-layer DNN classifier is trained on the oversampled data to capture nonlinear relationships in transaction patterns. Model learning is optimized using backpropagation with binary cross-entropy (BCE) loss and adaptive optimizers. To ensure robust and optimal selection of training settings, an extensive hyperparameter search is performed across hidden layers (2–4), neurons (32–128), activation functions (rectified linear unit (ReLU), Sigmoid, Softmax, Tanh), optimizers (Adam, root mean square propagation (RMSprop), stochastic gradient descent (SGD), learning rates (0.001–0.1), batch sizes (16–128), epochs (100–500), and loss functions (BCE, mean squared error (MSE), Hinge). Parallel training and early-stage pruning are used to maintain computational feasibility. Results demonstrate that the best-performing configuration is achieved with SMOTE and a DNN trained using 3 hidden layers, 128 neurons per layer, Tanh activation, RMSprop optimizer, learning rate 0.001, batch size 64, and 300 epochs with BCE loss. Across the top configurations, accuracy ranges from 93.98% to 99.99% with precision up to 0.98, recall up to 0.99, and F1-score up to 0.98. Overall, SMOTE outperforms ADASYN and borderline-SMOTE, achieving 98.92% accuracy, 0.95 precision, 0.96 recall, and a 0.95 F1-score. The results of receiver operating characteristic (ROC) curve, and area under the curve (AUC) (ROC–AUC) also confirm the same trend.
Keywords
Credit card fraud detection, Deep neural network (DNN), Class imbalance, SMOTE, ADASYN, Borderline-SMOTE.
Cite this article
Rinku, Dubey AK, Narang SK. An oversampling-integrated deep neural network for imbalanced credit card fraud detection. International Journal of Advanced Technology and Engineering Exploration. 2025;12(133):1852-1866. DOI : 10.19101/IJATEE.2024.111101641